The physical observer in a Szilard engine with uncertainty
Published in arXiv preprint arXiv:2309.10580, 2023
Recommended citation: Daimer, D. & Still, S. (2023). The physical observer in a Szilard engine with uncertainty. arXiv preprint arXiv:2309.10580.
We use the fact that an algorithm for computing optimal strategies can be directly derived from maximizing overall engine work output in generalized partially observable information engines. For a stylizedly simple decision problem, we discover interesting optimal strategies that differ notably from naive coarse graining. They inspire a model class of simple, yet compelling, parameterized soft partitionings.
We analyze and compare optimal strategies for three different observer classes: (1) optimal observers, (2) observers limited to the parameterized soft partitionings introduced here and (3) observers limited to coarse graining. While coarse graining based observers are outperformed by the other two types of observers, there is no difference in performance between unconstrained, optimal observers and those limited to soft partitionings. The parameterized soft partitioning strategies allow us to compute key quantities of the decision problem analytically.